By design, LLMs can get faster but cannot be more accurate without a massive intentional approach to verifying their datasets, which isn't feasible because that would counter anything not fact based as LLMs don't understand context. Basically, the training approach means that they get filled with whatever the builders can get their hands on and then they fall back to web searches which return all kinds of unreliable stuff because LLMs don't have a way of verifying reliability.

Even if they were perfect, they will not be able to keep up with the content flood of new information that comes out every minute when used as general purpose answer anything tools.

What AI actually excels at is pattern matching in controlled settings.

They're operating under the long outdated assumption that all you need to simulate a brain is match the number of neurons....

That's not how any of this works, but they've been saying "we'll be there soon" for so long now that we're almost able to do it, their gonna lose their main excuse and main reason for fundraising.

They'll have to tell investors the timeline just changed from years to maybe decades if we're lucky

And it's gonna divebomb our whole economy because fucking every fund manager is dumping insane levels of money into it.

I think the title should be "What if LLMs doesn't get much better than this?" because that's effectively what the article is talking about. I see no reason to expect that our AI systems wouldn't keep improving even if LLMs don't.

Some guy blogged that the smart ones move to advertising.

Honestly? If AI systems stopped improving forever? That's probably best case scenario. LLMs are already superhuman on a knowledge level, human-level in terms of speed (tokens per sec, etc), but subhuman in many other areas. This makes them useful for some tasks, but not so useful that they could cause any sort of existential threat to humanity (either in an economic sense or in a misalignment sense). If LLMs stagnate here then we have at least one tool in our AI toolbox that we're pretty sure isn't conscious/sentient/etc., which is useful since that makes them predictable on some level. Humans can deal with that.

Unfortunately, I see no reason why AI systems in general wouldn't continue to improve. Even if LLMs do stagnate they're only one tiny branch of a much larger tree, and we already have at least one example of an AI system that is conscious and sentient - a human. This means even if somehow the human brain was the only architecture ever capable of sentience (incredibly unlikely), we could always simulate/emulate a human brain to get human-level AGI. Simulate/emulate it faster? Superhuman AGI.

Next year: OpenAI declares bankruptcy and news articles will say, "what happened with AI?"

Industry collapses?